MMA-SafetyBench: A Benchmark for Multimodal Agent Safety Evaluation
- Published
- Source
- arXiv
- Paper number
- 1344
- Field
- AI Agents
- arXiv ID
- 2610.19384
Key points
- The authors name the failure mode in which plausible visual elements such as forged system alerts pull an agent away from its assigned task Visual Semantic Hijacking.
- They built MMA-SafetyBench with 1,083 adversarial trajectories spanning web automation, desktop GUI, mobile navigation, document understanding, and video processing.
- Across nine vision-language models, Attack Success Rate (ASR) reaches 87.96% and Semantic Compromise Rate (SCR) reaches 98.09%.
- A GPT-5 judge matched the human reference label on 290 of 300 cases (96.67%), supporting scoring reliability.
- With a Safety-Priority Prompt (SPP), average ASR over the 45 model-domain pairs falls from 55.20% to 19.54%, but residual failures remain substantial.
- As limitations, coordinate-hit ASR can mistake grounding errors for robustness, and the study does not measure attack prevalence in deployment.
Paper links
External research summaries. These are not HDATF publications or measured product results.